Deepsite: A Next-Generation Vibe Coding Platform for Browser-Based Application Generation

Abstract

DeepSite (v4) is a cloud-native, browser-based development platform designed to enable intuitive software creation through natural language interaction. With the growing capabilities of Large Language Models (LLMs), software development is evolving from manual coding toward intent-driven design. DeepSite addresses the challenges of this transition by providing a structured environment for generating, modifying, and executing AI-produced code efficiently. The platform integrates the Next.js App Router with Hugging Face serverless inference APIs to stream incremental code updates in real time. Instead of regenerating entire files, a constrained Search/Replace mechanism is employed to apply precise modifications, reducing token usage and minimizing inconsistencies. For execution, the system utilizes browser-based Web Containers (Sandpack), allowing instant preview without local setup. Project persistence is achieved through Hugging Face Spaces, enabling Git-based version control and decentralized storage without relying on traditional databases. The platform ensures scalability, security, and rapid prototyping by combining client-side execution with lightweight backend orchestration. This work demonstrates a practical approach to AI-assisted development and highlights the potential of browser-based environments in accelerating modern software engineering workflows.